A learning model for essentialist concepts

Many cognitive scientists take it for granted that concepts like CAT (mental terms that are expressed with single nouns) can be learned by observing a co-occurrence in superficial properties, such as having fur, being 4-legged, and tending to purr, and then building a complex category representation from representations for those superficial properties. A less popular account, known as Psychological Essentialism, claims that concepts like CAT pick out deep, hidden properties (essences) that are causal explanations for observable co-occurrences in superficial properties. The trouble is, Psychological Essentialism lacks an account of how such essentialist concepts could be learned, and often adopt the unpalatable conclusion that such concepts are innate. Developmental roboticists have recently started implementing systems that employ learned hidden/latent variables. The present paper spells out a learning theory for essentialist concepts, and presents two psychology experiments that help support the account over the associationist alternative.